Takes

Record once — the agent does the retakes. From any ready recording that carries an interaction trail, the agent reconstructs the flow you performed, re-performs it headlessly against a site you control, and records a fresh, watchable take: paced for a viewer, no stray cursor wanders, no typo at the worst moment. You review the draft and decide whether it becomes a recording at all.

The recording is the ground truth, not a suggestion. The agent performs flows it has seen you perform; it does not invent one from a text prompt. Your instruction can vary how the flow is performed — the route, the pacing, a toggle — but a take always starts from a real human recording.

What it’s for

• The clean retake. Your walkthrough was right but rough. One click gets a tidy version of the same flow, without performing it again yourself.
• The refresh after a UI change. The product moved on and the tutorial didn’t. Re-perform the old recording against the new build and re-analyse — a current guide from a stale one.
• The fix verification, on camera. On a bug report the same panel reads “Re-record against…”: after you ship the fix, the agent re-performs the report’s flow against your fixed build and records it working — evidence to close the issue with.
• The variation. “Do it slower on the settings step”, “use the sidebar instead of the menu”, “with the annual toggle on” — say it in words and the take follows them, grounded in the recorded flow.

How a take happens

1. Open a ready recording and choose Perform again. Pick the target origin (the recording’s own origin is prefilled), optionally tell the agent to do it differently, and run.
2. The agent plans the flow from the recording’s interaction log, then drives it in a real browser in our infrastructure, recording video and an exact log of every action it takes.
3. The take lands as a draft on the source recording’s page — visible only to you, not in the library, not shared, not analysed. Watch it, then approve it into a real recording (choosing the analysis mode) or discard it, which deletes the video immediately.
4. An approved take goes through the ordinary pipeline — analysis, chapters, exports, sharing — like any recording, with one permanent difference: provenance.

Provenance is not optional

An approved take is labelled performed by the agent everywhere it appears: the player page, share pages, embeds, and a provenance line in every export. The label is not removable on any plan — it is provenance, not branding. We never let machine output pass as human capture. Two consequences follow: a performed recording can’t be featured in the public gallery, and the Repro Agent refuses to re-drive one — re-verifying a performance with the same driver that performed it would be circular.

The same rules as the Repro Agent

A take drives real clicks and real requests at a real site, so the boundaries are identical to reproduction runs: only origins your workspace has verified, never private or internal addresses, never through a login — the agent has no credentials and will not type any. A flow that hits an auth wall ends there and the take says so. Runs happen in our infrastructure on a clean browser profile, not on anyone’s machine.

Allowances

Takes are metered from the same monthly agent run allowance as Repro Agent runs — one take is one run — and are available on Pro and above. Approving a take then uses a normal AI analysis from your plan’s existing quota, like any new recording.

What it is not good at

The honest list: flows behind a login, real payment or email/SMS steps, drag-and-drop and canvas-heavy interactions, flows that depend on precise timing or a specific device, and pages that have changed so much the recorded steps no longer exist — those takes end early and say why, rather than improvising. The agent also cannot narrate; pair an approved take with AI narration if you want a voiced version.